search

geaxgx / depthai_blazepose

MIT License
322 stars 68 forks source link

Pose estimation for multiple people #15

Open vipulkumar-developer opened 2 years ago

vipulkumar-developer commented 2 years ago

Hi, I'm using this repository to test the pose estimation on an OAK-IOT device for the edge-programming and not host-programming. It works perfectly for one person but it can't detect other people. Is there any setting or way to do this?

geaxgx commented 2 years ago

Hi, Unfortunately no, as Blazepose is a single pose model.

vipulkumar-developer commented 2 years ago

Do you think that people detection with Yolo can solve the problem? By using Yolo we could get the image cropped on the single person and then pass the various frame to MediaPipe.

geaxgx commented 2 years ago

Sure, if the people in the image are clearly separated in the image, we can apply the landmark regression model on each detected person, it will work fine. And no need to use yolo, the mediapipe pose detection model works well enough. The problem comes when there are 2 people close to each other, more precisely when a 2nd person appears in the boudning rotated rectangle around the first person detected. In this case, the landmark regression model gets confused and can associate landmarks (of the first person) with 2nd parson body parts. It is something you can try by yourself on this older repo: https://github.com/geaxgx/openvino_blazepose It runs blazepose on openvino so you can use it with a webcam for instance. There is a --multi_detection flag to force multiple person detection.